Deep Learning Model Identifies Mental Disorders in First-Visit Psychiatric Patients

A deep learning model designated TUCH-FT has been developed to perform multi-label classification of mental disorders among first-visit psychiatric patients using multisource clinical data. Reporting in Frontiers in Psychiatry, researchers designed the framework to address diagnostic difficulties caused by overlapping symptoms during initial clinical evaluations, developing the system using data from 5,802 first-visit psychiatric patients.

The predictive framework incorporates three distinct categories of patient information, drawing upon demographic variables, recorded clinical history, and item-level data from the Symptom Checklist-90. To establish reliable reference standards for model development and evaluation, final diagnostic labels were assigned through longitudinal follow-up, documented treatment response, and psychiatric expert evaluation.

The authors identified several key limitations in the standard FT-Transformer architecture, which conventionally compiles heterogeneous clinical variables into a single sequence. This standard approach can weaken source-specific structures, inadequately represent unclear clinical histories, and fail to model directional cross-source interactions as well as hierarchical comorbidity relationships between conditions.

To overcome these structural limitations, the investigators introduced four specific methodological components into the TUCH-FT architecture. The system combines a three-branch clinical feature encoding method for heterogeneous information, an uncertainty-aware embedding method for ambiguous clinical histories, a cross-group multi-head attention fusion mechanism to capture multisource interactions, and a hierarchical multi-label classification method to model comorbidity relationships.

In primary diagnostic performance assessments, TUCH-FT achieved a Macro-F1 score of 0.803 and a Micro-F1 score of 0.842. The model also recorded a Macro-average precision of 0.838, a Hamming loss of 0.095, and a label ranking average precision of 0.876.

Model reliability was subsequently assessed in a temporally separated validation cohort comprising 1,272 patients from the same centre. In this validation group, TUCH-FT demonstrated calibration performance with expected calibration errors ranging from 0.031 to 0.067, while Brier scores across categories ranged from a low of 0.072 for psychotic disorders to a high of 0.134 for anxiety disorders.

The evaluation also highlighted performance constraints, including an exact-match accuracy of 60.8% for complete diagnostic combinations, reflecting ongoing challenges in predicting co-occurring disorders. Despite these limitations in complex comorbidity prediction, the investigators concluded that TUCH-FT may provide a practical reference tool for differential clinical assessment among first-visit patients presenting with mood-related symptoms.

The research was conducted by a clinical team at Heilongjiang Provincial Hospital in Harbin, China, led by corresponding author Zhan Gao alongside Yu Zhang, Wenyu Ge, Zhiyu Cao, Hongyan Zhu, Yuchao Ma, Weiwei Wang, Yuqing Song, Yu Li, and Xiangna Zheng.

Subject of Research: Deep learning for early identification of mental disorders in first-visit psychiatric patients

Article Title: A multisource clinical deep learning model for early identification of mental disorders in first-visit psychiatric patients

Article References: Zhang, Y., Gao, Z., Ge, W., Cao, Z., Zhu, H., Ma, Y., Wang, W., Song, Y., Li, Y., & Zheng, X. (2026). A multisource clinical deep learning model for early identification of mental disorders in first-visit psychiatric patients. Frontiers in Psychiatry, 17, Article 1988768. https://doi.org/10.3389/fpsyt.2026.1988768

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1988768

Keywords: deep learning, computational psychiatry, mental disorders, first-visit psychiatric patients, multi-label classification, SCL-90, psychiatric diagnosis, clinical validation, multisource, clinical, deep, learning